# How AI agents send every trade confirmation within minutes

A brokerage must confirm each trade to each client individually after market close. One brokerage did this with 10 staff spending 2 hours a day reading a PDF ledger and typing messages. FetchSky replaced that with three AI agents: one extracts trades from the ledger using OCR and document AI, one matches each trade to the right client in the CRM, and one sends the confirmation through the official WhatsApp business API. The published result is 98% less manual effort, around 20 person-hours saved a day, no typing errors, every confirmation logged, and a cost of roughly $0.01 per message.

By Muhammad Ali Khoja, CSO. Published 2026-10-22.

## What was the problem with end-of-day trade confirmations?

At market close the brokerage's systems produced one PDF ledger listing every trade executed that day. Regulation required each trade to be confirmed to each client individually, not in a batch. Ten agents worked through the ledger by hand for about 2 hours every trading day, finding each client's trades, looking up their contact details and sending a message. The work was repetitive, time-pressured and error-prone, and a typing error in a confirmation is a compliance problem as well as a customer one.

## How do the three agents work together?

The automation is a pipeline of three narrow agents, each with one job and only the access that job needs.

1. **Extraction agent.** Reads the PDF ledger with OCR and document AI and produces structured rows: client reference, instrument, quantity, price, side, time. It has read access to the ledger and nothing else.
2. **Identity matching agent.** Takes each row and matches the client reference to a client record in the CRM to retrieve the verified contact number. It has read access to the CRM and cannot change records.
3. **Messaging agent.** Formats each confirmation and sends it through the official WhatsApp business API, then writes the delivery result to the log. It can send messages and cannot look up or alter client data.

Because each agent is narrow, a failure in one stage stops the pipeline at that stage rather than producing a wrong message. The whole run, from the ledger arriving to the last message sent, finishes within minutes.

## What made the automation safe enough for a regulated brokerage?

Three design choices did most of the work.

**Logging for compliance.** Every confirmation is logged: what was extracted, which client it was matched to, what was sent and whether it was delivered. FetchSky's general control is that everything the AI reads, decides and does is logged. For a brokerage, that log is the evidence that the regulatory duty was met for every trade, every day.

**The official WhatsApp business API.** Messages go through the official business API rather than an unofficial client or a personal number. That gives verified sender identity, delivery receipts and a channel the brokerage's compliance team can stand behind.

**Deterministic identity matching.** The identity step matches on CRM identifiers rather than asking a language model to guess which client a name refers to. Where a match cannot be made deterministically, the trade is held for a person rather than sent to the closest guess. FetchSky's control that a person reviews anything affecting a customer applies at exactly that point.

## What changed?

|  | Before | After |
| --- | --- | --- |
| People on the task | 10 agents | 3 AI agents, with exceptions reviewed by a person |
| Time per day | 2 hours each, about 20 person-hours | Around 20 person-hours saved a day |
| Manual effort | Full | 98% less |
| Typing errors | Present | None |
| Audit trail | Partial, by hand | Every confirmation logged |
| Cost per message | Staff time | Roughly $0.01 |

## Why does this pattern generalise?

The structure is extract, match, act. The same shape appears in FetchSky's other published automations: a monthly holdings report that reads 34 fund manager reports now runs in 3 minutes instead of 3–4 days, and a daily market newsletter that took an analyst 2–3 hours now takes 80% or more less time. Any workflow that starts with a document, needs a lookup against a system of record, and ends with a message or a file fits the pattern. The process for finding such a task and proving it in six weeks is described in [the AI in production guide](/insights/ai-in-production-guide).

To scope the same pipeline for your own end-of-day process, start with [AI transformation](/ai-transformation).

## Read next

- [Enterprise AI in production: the complete guide](https://fetchsky.com/insights/ai-in-production-guide)
- [AI transformation](https://fetchsky.com/ai-transformation)
- [What is agentic AI? A plain-English guide for operations leaders](https://fetchsky.com/insights/what-is-agentic-ai)

Source: https://fetchsky.com/insights/how-ai-agents-send-every-trade-confirmation-within-minutes